xref: /netbsd-src/external/apache2/llvm/dist/llvm/lib/Target/PowerPC/PPCMIPeephole.cpp (revision 82d56013d7b633d116a93943de88e08335357a7c)
1 //===-------------- PPCMIPeephole.cpp - MI Peephole Cleanups -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===---------------------------------------------------------------------===//
8 //
9 // This pass performs peephole optimizations to clean up ugly code
10 // sequences at the MachineInstruction layer.  It runs at the end of
11 // the SSA phases, following VSX swap removal.  A pass of dead code
12 // elimination follows this one for quick clean-up of any dead
13 // instructions introduced here.  Although we could do this as callbacks
14 // from the generic peephole pass, this would have a couple of bad
15 // effects:  it might remove optimization opportunities for VSX swap
16 // removal, and it would miss cleanups made possible following VSX
17 // swap removal.
18 //
19 //===---------------------------------------------------------------------===//
20 
21 #include "MCTargetDesc/PPCMCTargetDesc.h"
22 #include "MCTargetDesc/PPCPredicates.h"
23 #include "PPC.h"
24 #include "PPCInstrBuilder.h"
25 #include "PPCInstrInfo.h"
26 #include "PPCMachineFunctionInfo.h"
27 #include "PPCTargetMachine.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
30 #include "llvm/CodeGen/MachineDominators.h"
31 #include "llvm/CodeGen/MachineFunctionPass.h"
32 #include "llvm/CodeGen/MachineInstrBuilder.h"
33 #include "llvm/CodeGen/MachinePostDominators.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/InitializePasses.h"
36 #include "llvm/Support/Debug.h"
37 
38 using namespace llvm;
39 
40 #define DEBUG_TYPE "ppc-mi-peepholes"
41 
42 STATISTIC(RemoveTOCSave, "Number of TOC saves removed");
43 STATISTIC(MultiTOCSaves,
44           "Number of functions with multiple TOC saves that must be kept");
45 STATISTIC(NumTOCSavesInPrologue, "Number of TOC saves placed in the prologue");
46 STATISTIC(NumEliminatedSExt, "Number of eliminated sign-extensions");
47 STATISTIC(NumEliminatedZExt, "Number of eliminated zero-extensions");
48 STATISTIC(NumOptADDLIs, "Number of optimized ADD instruction fed by LI");
49 STATISTIC(NumConvertedToImmediateForm,
50           "Number of instructions converted to their immediate form");
51 STATISTIC(NumFunctionsEnteredInMIPeephole,
52           "Number of functions entered in PPC MI Peepholes");
53 STATISTIC(NumFixedPointIterations,
54           "Number of fixed-point iterations converting reg-reg instructions "
55           "to reg-imm ones");
56 STATISTIC(NumRotatesCollapsed,
57           "Number of pairs of rotate left, clear left/right collapsed");
58 STATISTIC(NumEXTSWAndSLDICombined,
59           "Number of pairs of EXTSW and SLDI combined as EXTSWSLI");
60 STATISTIC(NumLoadImmZeroFoldedAndRemoved,
61           "Number of LI(8) reg, 0 that are folded to r0 and removed");
62 
63 static cl::opt<bool>
64 FixedPointRegToImm("ppc-reg-to-imm-fixed-point", cl::Hidden, cl::init(true),
65                    cl::desc("Iterate to a fixed point when attempting to "
66                             "convert reg-reg instructions to reg-imm"));
67 
68 static cl::opt<bool>
69 ConvertRegReg("ppc-convert-rr-to-ri", cl::Hidden, cl::init(true),
70               cl::desc("Convert eligible reg+reg instructions to reg+imm"));
71 
72 static cl::opt<bool>
73     EnableSExtElimination("ppc-eliminate-signext",
74                           cl::desc("enable elimination of sign-extensions"),
75                           cl::init(false), cl::Hidden);
76 
77 static cl::opt<bool>
78     EnableZExtElimination("ppc-eliminate-zeroext",
79                           cl::desc("enable elimination of zero-extensions"),
80                           cl::init(false), cl::Hidden);
81 
82 namespace {
83 
84 struct PPCMIPeephole : public MachineFunctionPass {
85 
86   static char ID;
87   const PPCInstrInfo *TII;
88   MachineFunction *MF;
89   MachineRegisterInfo *MRI;
90 
PPCMIPeephole__anon52e8e82b0111::PPCMIPeephole91   PPCMIPeephole() : MachineFunctionPass(ID) {
92     initializePPCMIPeepholePass(*PassRegistry::getPassRegistry());
93   }
94 
95 private:
96   MachineDominatorTree *MDT;
97   MachinePostDominatorTree *MPDT;
98   MachineBlockFrequencyInfo *MBFI;
99   uint64_t EntryFreq;
100 
101   // Initialize class variables.
102   void initialize(MachineFunction &MFParm);
103 
104   // Perform peepholes.
105   bool simplifyCode(void);
106 
107   // Perform peepholes.
108   bool eliminateRedundantCompare(void);
109   bool eliminateRedundantTOCSaves(std::map<MachineInstr *, bool> &TOCSaves);
110   bool combineSEXTAndSHL(MachineInstr &MI, MachineInstr *&ToErase);
111   bool emitRLDICWhenLoweringJumpTables(MachineInstr &MI);
112   void UpdateTOCSaves(std::map<MachineInstr *, bool> &TOCSaves,
113                       MachineInstr *MI);
114 
115 public:
116 
getAnalysisUsage__anon52e8e82b0111::PPCMIPeephole117   void getAnalysisUsage(AnalysisUsage &AU) const override {
118     AU.addRequired<MachineDominatorTree>();
119     AU.addRequired<MachinePostDominatorTree>();
120     AU.addRequired<MachineBlockFrequencyInfo>();
121     AU.addPreserved<MachineDominatorTree>();
122     AU.addPreserved<MachinePostDominatorTree>();
123     AU.addPreserved<MachineBlockFrequencyInfo>();
124     MachineFunctionPass::getAnalysisUsage(AU);
125   }
126 
127   // Main entry point for this pass.
runOnMachineFunction__anon52e8e82b0111::PPCMIPeephole128   bool runOnMachineFunction(MachineFunction &MF) override {
129     initialize(MF);
130     // At this point, TOC pointer should not be used in a function that uses
131     // PC-Relative addressing.
132     assert((MF.getRegInfo().use_empty(PPC::X2) ||
133             !MF.getSubtarget<PPCSubtarget>().isUsingPCRelativeCalls()) &&
134            "TOC pointer used in a function using PC-Relative addressing!");
135     if (skipFunction(MF.getFunction()))
136       return false;
137     return simplifyCode();
138   }
139 };
140 
141 // Initialize class variables.
initialize(MachineFunction & MFParm)142 void PPCMIPeephole::initialize(MachineFunction &MFParm) {
143   MF = &MFParm;
144   MRI = &MF->getRegInfo();
145   MDT = &getAnalysis<MachineDominatorTree>();
146   MPDT = &getAnalysis<MachinePostDominatorTree>();
147   MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
148   EntryFreq = MBFI->getEntryFreq();
149   TII = MF->getSubtarget<PPCSubtarget>().getInstrInfo();
150   LLVM_DEBUG(dbgs() << "*** PowerPC MI peephole pass ***\n\n");
151   LLVM_DEBUG(MF->dump());
152 }
153 
getVRegDefOrNull(MachineOperand * Op,MachineRegisterInfo * MRI)154 static MachineInstr *getVRegDefOrNull(MachineOperand *Op,
155                                       MachineRegisterInfo *MRI) {
156   assert(Op && "Invalid Operand!");
157   if (!Op->isReg())
158     return nullptr;
159 
160   Register Reg = Op->getReg();
161   if (!Register::isVirtualRegister(Reg))
162     return nullptr;
163 
164   return MRI->getVRegDef(Reg);
165 }
166 
167 // This function returns number of known zero bits in output of MI
168 // starting from the most significant bit.
169 static unsigned
getKnownLeadingZeroCount(MachineInstr * MI,const PPCInstrInfo * TII)170 getKnownLeadingZeroCount(MachineInstr *MI, const PPCInstrInfo *TII) {
171   unsigned Opcode = MI->getOpcode();
172   if (Opcode == PPC::RLDICL || Opcode == PPC::RLDICL_rec ||
173       Opcode == PPC::RLDCL || Opcode == PPC::RLDCL_rec)
174     return MI->getOperand(3).getImm();
175 
176   if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDIC_rec) &&
177       MI->getOperand(3).getImm() <= 63 - MI->getOperand(2).getImm())
178     return MI->getOperand(3).getImm();
179 
180   if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec ||
181        Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec ||
182        Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) &&
183       MI->getOperand(3).getImm() <= MI->getOperand(4).getImm())
184     return 32 + MI->getOperand(3).getImm();
185 
186   if (Opcode == PPC::ANDI_rec) {
187     uint16_t Imm = MI->getOperand(2).getImm();
188     return 48 + countLeadingZeros(Imm);
189   }
190 
191   if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZW_rec ||
192       Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZW_rec ||
193       Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8)
194     // The result ranges from 0 to 32.
195     return 58;
196 
197   if (Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZD_rec ||
198       Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZD_rec)
199     // The result ranges from 0 to 64.
200     return 57;
201 
202   if (Opcode == PPC::LHZ   || Opcode == PPC::LHZX  ||
203       Opcode == PPC::LHZ8  || Opcode == PPC::LHZX8 ||
204       Opcode == PPC::LHZU  || Opcode == PPC::LHZUX ||
205       Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8)
206     return 48;
207 
208   if (Opcode == PPC::LBZ   || Opcode == PPC::LBZX  ||
209       Opcode == PPC::LBZ8  || Opcode == PPC::LBZX8 ||
210       Opcode == PPC::LBZU  || Opcode == PPC::LBZUX ||
211       Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8)
212     return 56;
213 
214   if (TII->isZeroExtended(*MI))
215     return 32;
216 
217   return 0;
218 }
219 
220 // This function maintains a map for the pairs <TOC Save Instr, Keep>
221 // Each time a new TOC save is encountered, it checks if any of the existing
222 // ones are dominated by the new one. If so, it marks the existing one as
223 // redundant by setting it's entry in the map as false. It then adds the new
224 // instruction to the map with either true or false depending on if any
225 // existing instructions dominated the new one.
UpdateTOCSaves(std::map<MachineInstr *,bool> & TOCSaves,MachineInstr * MI)226 void PPCMIPeephole::UpdateTOCSaves(
227   std::map<MachineInstr *, bool> &TOCSaves, MachineInstr *MI) {
228   assert(TII->isTOCSaveMI(*MI) && "Expecting a TOC save instruction here");
229   // FIXME: Saving TOC in prologue hasn't been implemented well in AIX ABI part,
230   // here only support it under ELFv2.
231   if (MF->getSubtarget<PPCSubtarget>().isELFv2ABI()) {
232     PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>();
233 
234     MachineBasicBlock *Entry = &MF->front();
235     uint64_t CurrBlockFreq = MBFI->getBlockFreq(MI->getParent()).getFrequency();
236 
237     // If the block in which the TOC save resides is in a block that
238     // post-dominates Entry, or a block that is hotter than entry (keep in mind
239     // that early MachineLICM has already run so the TOC save won't be hoisted)
240     // we can just do the save in the prologue.
241     if (CurrBlockFreq > EntryFreq || MPDT->dominates(MI->getParent(), Entry))
242       FI->setMustSaveTOC(true);
243 
244     // If we are saving the TOC in the prologue, all the TOC saves can be
245     // removed from the code.
246     if (FI->mustSaveTOC()) {
247       for (auto &TOCSave : TOCSaves)
248         TOCSave.second = false;
249       // Add new instruction to map.
250       TOCSaves[MI] = false;
251       return;
252     }
253   }
254 
255   bool Keep = true;
256   for (auto It = TOCSaves.begin(); It != TOCSaves.end(); It++ ) {
257     MachineInstr *CurrInst = It->first;
258     // If new instruction dominates an existing one, mark existing one as
259     // redundant.
260     if (It->second && MDT->dominates(MI, CurrInst))
261       It->second = false;
262     // Check if the new instruction is redundant.
263     if (MDT->dominates(CurrInst, MI)) {
264       Keep = false;
265       break;
266     }
267   }
268   // Add new instruction to map.
269   TOCSaves[MI] = Keep;
270 }
271 
272 // This function returns a list of all PHI nodes in the tree starting from
273 // the RootPHI node. We perform a BFS traversal to get an ordered list of nodes.
274 // The list initially only contains the root PHI. When we visit a PHI node, we
275 // add it to the list. We continue to look for other PHI node operands while
276 // there are nodes to visit in the list. The function returns false if the
277 // optimization cannot be applied on this tree.
collectUnprimedAccPHIs(MachineRegisterInfo * MRI,MachineInstr * RootPHI,SmallVectorImpl<MachineInstr * > & PHIs)278 static bool collectUnprimedAccPHIs(MachineRegisterInfo *MRI,
279                                    MachineInstr *RootPHI,
280                                    SmallVectorImpl<MachineInstr *> &PHIs) {
281   PHIs.push_back(RootPHI);
282   unsigned VisitedIndex = 0;
283   while (VisitedIndex < PHIs.size()) {
284     MachineInstr *VisitedPHI = PHIs[VisitedIndex];
285     for (unsigned PHIOp = 1, NumOps = VisitedPHI->getNumOperands();
286          PHIOp != NumOps; PHIOp += 2) {
287       Register RegOp = VisitedPHI->getOperand(PHIOp).getReg();
288       if (!Register::isVirtualRegister(RegOp))
289         return false;
290       MachineInstr *Instr = MRI->getVRegDef(RegOp);
291       // While collecting the PHI nodes, we check if they can be converted (i.e.
292       // all the operands are either copies, implicit defs or PHI nodes).
293       unsigned Opcode = Instr->getOpcode();
294       if (Opcode == PPC::COPY) {
295         Register Reg = Instr->getOperand(1).getReg();
296         if (!Register::isVirtualRegister(Reg) ||
297             MRI->getRegClass(Reg) != &PPC::ACCRCRegClass)
298           return false;
299       } else if (Opcode != PPC::IMPLICIT_DEF && Opcode != PPC::PHI)
300         return false;
301       // If we detect a cycle in the PHI nodes, we exit. It would be
302       // possible to change cycles as well, but that would add a lot
303       // of complexity for a case that is unlikely to occur with MMA
304       // code.
305       if (Opcode != PPC::PHI)
306         continue;
307       if (llvm::is_contained(PHIs, Instr))
308         return false;
309       PHIs.push_back(Instr);
310     }
311     VisitedIndex++;
312   }
313   return true;
314 }
315 
316 // This function changes the unprimed accumulator PHI nodes in the PHIs list to
317 // primed accumulator PHI nodes. The list is traversed in reverse order to
318 // change all the PHI operands of a PHI node before changing the node itself.
319 // We keep a map to associate each changed PHI node to its non-changed form.
convertUnprimedAccPHIs(const PPCInstrInfo * TII,MachineRegisterInfo * MRI,SmallVectorImpl<MachineInstr * > & PHIs,Register Dst)320 static void convertUnprimedAccPHIs(const PPCInstrInfo *TII,
321                                    MachineRegisterInfo *MRI,
322                                    SmallVectorImpl<MachineInstr *> &PHIs,
323                                    Register Dst) {
324   DenseMap<MachineInstr *, MachineInstr *> ChangedPHIMap;
325   for (auto It = PHIs.rbegin(), End = PHIs.rend(); It != End; ++It) {
326     MachineInstr *PHI = *It;
327     SmallVector<std::pair<MachineOperand, MachineOperand>, 4> PHIOps;
328     // We check if the current PHI node can be changed by looking at its
329     // operands. If all the operands are either copies from primed
330     // accumulators, implicit definitions or other unprimed accumulator
331     // PHI nodes, we change it.
332     for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps;
333          PHIOp += 2) {
334       Register RegOp = PHI->getOperand(PHIOp).getReg();
335       MachineInstr *PHIInput = MRI->getVRegDef(RegOp);
336       unsigned Opcode = PHIInput->getOpcode();
337       assert((Opcode == PPC::COPY || Opcode == PPC::IMPLICIT_DEF ||
338               Opcode == PPC::PHI) &&
339              "Unexpected instruction");
340       if (Opcode == PPC::COPY) {
341         assert(MRI->getRegClass(PHIInput->getOperand(1).getReg()) ==
342                    &PPC::ACCRCRegClass &&
343                "Unexpected register class");
344         PHIOps.push_back({PHIInput->getOperand(1), PHI->getOperand(PHIOp + 1)});
345       } else if (Opcode == PPC::IMPLICIT_DEF) {
346         Register AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass);
347         BuildMI(*PHIInput->getParent(), PHIInput, PHIInput->getDebugLoc(),
348                 TII->get(PPC::IMPLICIT_DEF), AccReg);
349         PHIOps.push_back({MachineOperand::CreateReg(AccReg, false),
350                           PHI->getOperand(PHIOp + 1)});
351       } else if (Opcode == PPC::PHI) {
352         // We found a PHI operand. At this point we know this operand
353         // has already been changed so we get its associated changed form
354         // from the map.
355         assert(ChangedPHIMap.count(PHIInput) == 1 &&
356                "This PHI node should have already been changed.");
357         MachineInstr *PrimedAccPHI = ChangedPHIMap.lookup(PHIInput);
358         PHIOps.push_back({MachineOperand::CreateReg(
359                               PrimedAccPHI->getOperand(0).getReg(), false),
360                           PHI->getOperand(PHIOp + 1)});
361       }
362     }
363     Register AccReg = Dst;
364     // If the PHI node we are changing is the root node, the register it defines
365     // will be the destination register of the original copy (of the PHI def).
366     // For all other PHI's in the list, we need to create another primed
367     // accumulator virtual register as the PHI will no longer define the
368     // unprimed accumulator.
369     if (PHI != PHIs[0])
370       AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass);
371     MachineInstrBuilder NewPHI = BuildMI(
372         *PHI->getParent(), PHI, PHI->getDebugLoc(), TII->get(PPC::PHI), AccReg);
373     for (auto RegMBB : PHIOps)
374       NewPHI.add(RegMBB.first).add(RegMBB.second);
375     ChangedPHIMap[PHI] = NewPHI.getInstr();
376   }
377 }
378 
379 // Perform peephole optimizations.
simplifyCode(void)380 bool PPCMIPeephole::simplifyCode(void) {
381   bool Simplified = false;
382   MachineInstr* ToErase = nullptr;
383   std::map<MachineInstr *, bool> TOCSaves;
384   const TargetRegisterInfo *TRI = &TII->getRegisterInfo();
385   NumFunctionsEnteredInMIPeephole++;
386   if (ConvertRegReg) {
387     // Fixed-point conversion of reg/reg instructions fed by load-immediate
388     // into reg/imm instructions. FIXME: This is expensive, control it with
389     // an option.
390     bool SomethingChanged = false;
391     do {
392       NumFixedPointIterations++;
393       SomethingChanged = false;
394       for (MachineBasicBlock &MBB : *MF) {
395         for (MachineInstr &MI : MBB) {
396           if (MI.isDebugInstr())
397             continue;
398 
399           if (TII->convertToImmediateForm(MI)) {
400             // We don't erase anything in case the def has other uses. Let DCE
401             // remove it if it can be removed.
402             LLVM_DEBUG(dbgs() << "Converted instruction to imm form: ");
403             LLVM_DEBUG(MI.dump());
404             NumConvertedToImmediateForm++;
405             SomethingChanged = true;
406             Simplified = true;
407             continue;
408           }
409         }
410       }
411     } while (SomethingChanged && FixedPointRegToImm);
412   }
413 
414   for (MachineBasicBlock &MBB : *MF) {
415     for (MachineInstr &MI : MBB) {
416 
417       // If the previous instruction was marked for elimination,
418       // remove it now.
419       if (ToErase) {
420         ToErase->eraseFromParent();
421         ToErase = nullptr;
422       }
423 
424       // Ignore debug instructions.
425       if (MI.isDebugInstr())
426         continue;
427 
428       // Per-opcode peepholes.
429       switch (MI.getOpcode()) {
430 
431       default:
432         break;
433       case PPC::COPY: {
434         Register Src = MI.getOperand(1).getReg();
435         Register Dst = MI.getOperand(0).getReg();
436         if (!Register::isVirtualRegister(Src) ||
437             !Register::isVirtualRegister(Dst))
438           break;
439         if (MRI->getRegClass(Src) != &PPC::UACCRCRegClass ||
440             MRI->getRegClass(Dst) != &PPC::ACCRCRegClass)
441           break;
442 
443         // We are copying an unprimed accumulator to a primed accumulator.
444         // If the input to the copy is a PHI that is fed only by (i) copies in
445         // the other direction (ii) implicitly defined unprimed accumulators or
446         // (iii) other PHI nodes satisfying (i) and (ii), we can change
447         // the PHI to a PHI on primed accumulators (as long as we also change
448         // its operands). To detect and change such copies, we first get a list
449         // of all the PHI nodes starting from the root PHI node in BFS order.
450         // We then visit all these PHI nodes to check if they can be changed to
451         // primed accumulator PHI nodes and if so, we change them.
452         MachineInstr *RootPHI = MRI->getVRegDef(Src);
453         if (RootPHI->getOpcode() != PPC::PHI)
454           break;
455 
456         SmallVector<MachineInstr *, 4> PHIs;
457         if (!collectUnprimedAccPHIs(MRI, RootPHI, PHIs))
458           break;
459 
460         convertUnprimedAccPHIs(TII, MRI, PHIs, Dst);
461 
462         ToErase = &MI;
463         break;
464       }
465       case PPC::LI:
466       case PPC::LI8: {
467         // If we are materializing a zero, look for any use operands for which
468         // zero means immediate zero. All such operands can be replaced with
469         // PPC::ZERO.
470         if (!MI.getOperand(1).isImm() || MI.getOperand(1).getImm() != 0)
471           break;
472         unsigned MIDestReg = MI.getOperand(0).getReg();
473         for (MachineInstr& UseMI : MRI->use_instructions(MIDestReg))
474           Simplified |= TII->onlyFoldImmediate(UseMI, MI, MIDestReg);
475         if (MRI->use_nodbg_empty(MIDestReg)) {
476           ++NumLoadImmZeroFoldedAndRemoved;
477           ToErase = &MI;
478         }
479         break;
480       }
481       case PPC::STW:
482       case PPC::STD: {
483         MachineFrameInfo &MFI = MF->getFrameInfo();
484         if (MFI.hasVarSizedObjects() ||
485             (!MF->getSubtarget<PPCSubtarget>().isELFv2ABI() &&
486              !MF->getSubtarget<PPCSubtarget>().isAIXABI()))
487           break;
488         // When encountering a TOC save instruction, call UpdateTOCSaves
489         // to add it to the TOCSaves map and mark any existing TOC saves
490         // it dominates as redundant.
491         if (TII->isTOCSaveMI(MI))
492           UpdateTOCSaves(TOCSaves, &MI);
493         break;
494       }
495       case PPC::XXPERMDI: {
496         // Perform simplifications of 2x64 vector swaps and splats.
497         // A swap is identified by an immediate value of 2, and a splat
498         // is identified by an immediate value of 0 or 3.
499         int Immed = MI.getOperand(3).getImm();
500 
501         if (Immed == 1)
502           break;
503 
504         // For each of these simplifications, we need the two source
505         // regs to match.  Unfortunately, MachineCSE ignores COPY and
506         // SUBREG_TO_REG, so for example we can see
507         //   XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), immed.
508         // We have to look through chains of COPY and SUBREG_TO_REG
509         // to find the real source values for comparison.
510         unsigned TrueReg1 =
511           TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI);
512         unsigned TrueReg2 =
513           TRI->lookThruCopyLike(MI.getOperand(2).getReg(), MRI);
514 
515         if (!(TrueReg1 == TrueReg2 && Register::isVirtualRegister(TrueReg1)))
516           break;
517 
518         MachineInstr *DefMI = MRI->getVRegDef(TrueReg1);
519 
520         if (!DefMI)
521           break;
522 
523         unsigned DefOpc = DefMI->getOpcode();
524 
525         // If this is a splat fed by a splatting load, the splat is
526         // redundant. Replace with a copy. This doesn't happen directly due
527         // to code in PPCDAGToDAGISel.cpp, but it can happen when converting
528         // a load of a double to a vector of 64-bit integers.
529         auto isConversionOfLoadAndSplat = [=]() -> bool {
530           if (DefOpc != PPC::XVCVDPSXDS && DefOpc != PPC::XVCVDPUXDS)
531             return false;
532           unsigned FeedReg1 =
533             TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI);
534           if (Register::isVirtualRegister(FeedReg1)) {
535             MachineInstr *LoadMI = MRI->getVRegDef(FeedReg1);
536             if (LoadMI && LoadMI->getOpcode() == PPC::LXVDSX)
537               return true;
538           }
539           return false;
540         };
541         if ((Immed == 0 || Immed == 3) &&
542             (DefOpc == PPC::LXVDSX || isConversionOfLoadAndSplat())) {
543           LLVM_DEBUG(dbgs() << "Optimizing load-and-splat/splat "
544                                "to load-and-splat/copy: ");
545           LLVM_DEBUG(MI.dump());
546           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
547                   MI.getOperand(0).getReg())
548               .add(MI.getOperand(1));
549           ToErase = &MI;
550           Simplified = true;
551         }
552 
553         // If this is a splat or a swap fed by another splat, we
554         // can replace it with a copy.
555         if (DefOpc == PPC::XXPERMDI) {
556           unsigned DefReg1 = DefMI->getOperand(1).getReg();
557           unsigned DefReg2 = DefMI->getOperand(2).getReg();
558           unsigned DefImmed = DefMI->getOperand(3).getImm();
559 
560           // If the two inputs are not the same register, check to see if
561           // they originate from the same virtual register after only
562           // copy-like instructions.
563           if (DefReg1 != DefReg2) {
564             unsigned FeedReg1 = TRI->lookThruCopyLike(DefReg1, MRI);
565             unsigned FeedReg2 = TRI->lookThruCopyLike(DefReg2, MRI);
566 
567             if (!(FeedReg1 == FeedReg2 &&
568                   Register::isVirtualRegister(FeedReg1)))
569               break;
570           }
571 
572           if (DefImmed == 0 || DefImmed == 3) {
573             LLVM_DEBUG(dbgs() << "Optimizing splat/swap or splat/splat "
574                                  "to splat/copy: ");
575             LLVM_DEBUG(MI.dump());
576             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
577                     MI.getOperand(0).getReg())
578                 .add(MI.getOperand(1));
579             ToErase = &MI;
580             Simplified = true;
581           }
582 
583           // If this is a splat fed by a swap, we can simplify modify
584           // the splat to splat the other value from the swap's input
585           // parameter.
586           else if ((Immed == 0 || Immed == 3) && DefImmed == 2) {
587             LLVM_DEBUG(dbgs() << "Optimizing swap/splat => splat: ");
588             LLVM_DEBUG(MI.dump());
589             MI.getOperand(1).setReg(DefReg1);
590             MI.getOperand(2).setReg(DefReg2);
591             MI.getOperand(3).setImm(3 - Immed);
592             Simplified = true;
593           }
594 
595           // If this is a swap fed by a swap, we can replace it
596           // with a copy from the first swap's input.
597           else if (Immed == 2 && DefImmed == 2) {
598             LLVM_DEBUG(dbgs() << "Optimizing swap/swap => copy: ");
599             LLVM_DEBUG(MI.dump());
600             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
601                     MI.getOperand(0).getReg())
602                 .add(DefMI->getOperand(1));
603             ToErase = &MI;
604             Simplified = true;
605           }
606         } else if ((Immed == 0 || Immed == 3) && DefOpc == PPC::XXPERMDIs &&
607                    (DefMI->getOperand(2).getImm() == 0 ||
608                     DefMI->getOperand(2).getImm() == 3)) {
609           // Splat fed by another splat - switch the output of the first
610           // and remove the second.
611           DefMI->getOperand(0).setReg(MI.getOperand(0).getReg());
612           ToErase = &MI;
613           Simplified = true;
614           LLVM_DEBUG(dbgs() << "Removing redundant splat: ");
615           LLVM_DEBUG(MI.dump());
616         }
617         break;
618       }
619       case PPC::VSPLTB:
620       case PPC::VSPLTH:
621       case PPC::XXSPLTW: {
622         unsigned MyOpcode = MI.getOpcode();
623         unsigned OpNo = MyOpcode == PPC::XXSPLTW ? 1 : 2;
624         unsigned TrueReg =
625           TRI->lookThruCopyLike(MI.getOperand(OpNo).getReg(), MRI);
626         if (!Register::isVirtualRegister(TrueReg))
627           break;
628         MachineInstr *DefMI = MRI->getVRegDef(TrueReg);
629         if (!DefMI)
630           break;
631         unsigned DefOpcode = DefMI->getOpcode();
632         auto isConvertOfSplat = [=]() -> bool {
633           if (DefOpcode != PPC::XVCVSPSXWS && DefOpcode != PPC::XVCVSPUXWS)
634             return false;
635           Register ConvReg = DefMI->getOperand(1).getReg();
636           if (!Register::isVirtualRegister(ConvReg))
637             return false;
638           MachineInstr *Splt = MRI->getVRegDef(ConvReg);
639           return Splt && (Splt->getOpcode() == PPC::LXVWSX ||
640             Splt->getOpcode() == PPC::XXSPLTW);
641         };
642         bool AlreadySplat = (MyOpcode == DefOpcode) ||
643           (MyOpcode == PPC::VSPLTB && DefOpcode == PPC::VSPLTBs) ||
644           (MyOpcode == PPC::VSPLTH && DefOpcode == PPC::VSPLTHs) ||
645           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::XXSPLTWs) ||
646           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::LXVWSX) ||
647           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::MTVSRWS)||
648           (MyOpcode == PPC::XXSPLTW && isConvertOfSplat());
649         // If the instruction[s] that feed this splat have already splat
650         // the value, this splat is redundant.
651         if (AlreadySplat) {
652           LLVM_DEBUG(dbgs() << "Changing redundant splat to a copy: ");
653           LLVM_DEBUG(MI.dump());
654           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
655                   MI.getOperand(0).getReg())
656               .add(MI.getOperand(OpNo));
657           ToErase = &MI;
658           Simplified = true;
659         }
660         // Splat fed by a shift. Usually when we align value to splat into
661         // vector element zero.
662         if (DefOpcode == PPC::XXSLDWI) {
663           Register ShiftRes = DefMI->getOperand(0).getReg();
664           Register ShiftOp1 = DefMI->getOperand(1).getReg();
665           Register ShiftOp2 = DefMI->getOperand(2).getReg();
666           unsigned ShiftImm = DefMI->getOperand(3).getImm();
667           unsigned SplatImm = MI.getOperand(2).getImm();
668           if (ShiftOp1 == ShiftOp2) {
669             unsigned NewElem = (SplatImm + ShiftImm) & 0x3;
670             if (MRI->hasOneNonDBGUse(ShiftRes)) {
671               LLVM_DEBUG(dbgs() << "Removing redundant shift: ");
672               LLVM_DEBUG(DefMI->dump());
673               ToErase = DefMI;
674             }
675             Simplified = true;
676             LLVM_DEBUG(dbgs() << "Changing splat immediate from " << SplatImm
677                               << " to " << NewElem << " in instruction: ");
678             LLVM_DEBUG(MI.dump());
679             MI.getOperand(1).setReg(ShiftOp1);
680             MI.getOperand(2).setImm(NewElem);
681           }
682         }
683         break;
684       }
685       case PPC::XVCVDPSP: {
686         // If this is a DP->SP conversion fed by an FRSP, the FRSP is redundant.
687         unsigned TrueReg =
688           TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI);
689         if (!Register::isVirtualRegister(TrueReg))
690           break;
691         MachineInstr *DefMI = MRI->getVRegDef(TrueReg);
692 
693         // This can occur when building a vector of single precision or integer
694         // values.
695         if (DefMI && DefMI->getOpcode() == PPC::XXPERMDI) {
696           unsigned DefsReg1 =
697             TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI);
698           unsigned DefsReg2 =
699             TRI->lookThruCopyLike(DefMI->getOperand(2).getReg(), MRI);
700           if (!Register::isVirtualRegister(DefsReg1) ||
701               !Register::isVirtualRegister(DefsReg2))
702             break;
703           MachineInstr *P1 = MRI->getVRegDef(DefsReg1);
704           MachineInstr *P2 = MRI->getVRegDef(DefsReg2);
705 
706           if (!P1 || !P2)
707             break;
708 
709           // Remove the passed FRSP/XSRSP instruction if it only feeds this MI
710           // and set any uses of that FRSP/XSRSP (in this MI) to the source of
711           // the FRSP/XSRSP.
712           auto removeFRSPIfPossible = [&](MachineInstr *RoundInstr) {
713             unsigned Opc = RoundInstr->getOpcode();
714             if ((Opc == PPC::FRSP || Opc == PPC::XSRSP) &&
715                 MRI->hasOneNonDBGUse(RoundInstr->getOperand(0).getReg())) {
716               Simplified = true;
717               Register ConvReg1 = RoundInstr->getOperand(1).getReg();
718               Register FRSPDefines = RoundInstr->getOperand(0).getReg();
719               MachineInstr &Use = *(MRI->use_instr_nodbg_begin(FRSPDefines));
720               for (int i = 0, e = Use.getNumOperands(); i < e; ++i)
721                 if (Use.getOperand(i).isReg() &&
722                     Use.getOperand(i).getReg() == FRSPDefines)
723                   Use.getOperand(i).setReg(ConvReg1);
724               LLVM_DEBUG(dbgs() << "Removing redundant FRSP/XSRSP:\n");
725               LLVM_DEBUG(RoundInstr->dump());
726               LLVM_DEBUG(dbgs() << "As it feeds instruction:\n");
727               LLVM_DEBUG(MI.dump());
728               LLVM_DEBUG(dbgs() << "Through instruction:\n");
729               LLVM_DEBUG(DefMI->dump());
730               RoundInstr->eraseFromParent();
731             }
732           };
733 
734           // If the input to XVCVDPSP is a vector that was built (even
735           // partially) out of FRSP's, the FRSP(s) can safely be removed
736           // since this instruction performs the same operation.
737           if (P1 != P2) {
738             removeFRSPIfPossible(P1);
739             removeFRSPIfPossible(P2);
740             break;
741           }
742           removeFRSPIfPossible(P1);
743         }
744         break;
745       }
746       case PPC::EXTSH:
747       case PPC::EXTSH8:
748       case PPC::EXTSH8_32_64: {
749         if (!EnableSExtElimination) break;
750         Register NarrowReg = MI.getOperand(1).getReg();
751         if (!Register::isVirtualRegister(NarrowReg))
752           break;
753 
754         MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg);
755         // If we've used a zero-extending load that we will sign-extend,
756         // just do a sign-extending load.
757         if (SrcMI->getOpcode() == PPC::LHZ ||
758             SrcMI->getOpcode() == PPC::LHZX) {
759           if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg()))
760             break;
761           auto is64Bit = [] (unsigned Opcode) {
762             return Opcode == PPC::EXTSH8;
763           };
764           auto isXForm = [] (unsigned Opcode) {
765             return Opcode == PPC::LHZX;
766           };
767           auto getSextLoadOp = [] (bool is64Bit, bool isXForm) {
768             if (is64Bit)
769               if (isXForm) return PPC::LHAX8;
770               else         return PPC::LHA8;
771             else
772               if (isXForm) return PPC::LHAX;
773               else         return PPC::LHA;
774           };
775           unsigned Opc = getSextLoadOp(is64Bit(MI.getOpcode()),
776                                        isXForm(SrcMI->getOpcode()));
777           LLVM_DEBUG(dbgs() << "Zero-extending load\n");
778           LLVM_DEBUG(SrcMI->dump());
779           LLVM_DEBUG(dbgs() << "and sign-extension\n");
780           LLVM_DEBUG(MI.dump());
781           LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n");
782           SrcMI->setDesc(TII->get(Opc));
783           SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg());
784           ToErase = &MI;
785           Simplified = true;
786           NumEliminatedSExt++;
787         }
788         break;
789       }
790       case PPC::EXTSW:
791       case PPC::EXTSW_32:
792       case PPC::EXTSW_32_64: {
793         if (!EnableSExtElimination) break;
794         Register NarrowReg = MI.getOperand(1).getReg();
795         if (!Register::isVirtualRegister(NarrowReg))
796           break;
797 
798         MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg);
799         // If we've used a zero-extending load that we will sign-extend,
800         // just do a sign-extending load.
801         if (SrcMI->getOpcode() == PPC::LWZ ||
802             SrcMI->getOpcode() == PPC::LWZX) {
803           if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg()))
804             break;
805           auto is64Bit = [] (unsigned Opcode) {
806             return Opcode == PPC::EXTSW || Opcode == PPC::EXTSW_32_64;
807           };
808           auto isXForm = [] (unsigned Opcode) {
809             return Opcode == PPC::LWZX;
810           };
811           auto getSextLoadOp = [] (bool is64Bit, bool isXForm) {
812             if (is64Bit)
813               if (isXForm) return PPC::LWAX;
814               else         return PPC::LWA;
815             else
816               if (isXForm) return PPC::LWAX_32;
817               else         return PPC::LWA_32;
818           };
819           unsigned Opc = getSextLoadOp(is64Bit(MI.getOpcode()),
820                                        isXForm(SrcMI->getOpcode()));
821           LLVM_DEBUG(dbgs() << "Zero-extending load\n");
822           LLVM_DEBUG(SrcMI->dump());
823           LLVM_DEBUG(dbgs() << "and sign-extension\n");
824           LLVM_DEBUG(MI.dump());
825           LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n");
826           SrcMI->setDesc(TII->get(Opc));
827           SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg());
828           ToErase = &MI;
829           Simplified = true;
830           NumEliminatedSExt++;
831         } else if (MI.getOpcode() == PPC::EXTSW_32_64 &&
832                    TII->isSignExtended(*SrcMI)) {
833           // We can eliminate EXTSW if the input is known to be already
834           // sign-extended.
835           LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n");
836           Register TmpReg =
837               MF->getRegInfo().createVirtualRegister(&PPC::G8RCRegClass);
838           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::IMPLICIT_DEF),
839                   TmpReg);
840           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::INSERT_SUBREG),
841                   MI.getOperand(0).getReg())
842               .addReg(TmpReg)
843               .addReg(NarrowReg)
844               .addImm(PPC::sub_32);
845           ToErase = &MI;
846           Simplified = true;
847           NumEliminatedSExt++;
848         }
849         break;
850       }
851       case PPC::RLDICL: {
852         // We can eliminate RLDICL (e.g. for zero-extension)
853         // if all bits to clear are already zero in the input.
854         // This code assume following code sequence for zero-extension.
855         //   %6 = COPY %5:sub_32; (optional)
856         //   %8 = IMPLICIT_DEF;
857         //   %7<def,tied1> = INSERT_SUBREG %8<tied0>, %6, sub_32;
858         if (!EnableZExtElimination) break;
859 
860         if (MI.getOperand(2).getImm() != 0)
861           break;
862 
863         Register SrcReg = MI.getOperand(1).getReg();
864         if (!Register::isVirtualRegister(SrcReg))
865           break;
866 
867         MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
868         if (!(SrcMI && SrcMI->getOpcode() == PPC::INSERT_SUBREG &&
869               SrcMI->getOperand(0).isReg() && SrcMI->getOperand(1).isReg()))
870           break;
871 
872         MachineInstr *ImpDefMI, *SubRegMI;
873         ImpDefMI = MRI->getVRegDef(SrcMI->getOperand(1).getReg());
874         SubRegMI = MRI->getVRegDef(SrcMI->getOperand(2).getReg());
875         if (ImpDefMI->getOpcode() != PPC::IMPLICIT_DEF) break;
876 
877         SrcMI = SubRegMI;
878         if (SubRegMI->getOpcode() == PPC::COPY) {
879           Register CopyReg = SubRegMI->getOperand(1).getReg();
880           if (Register::isVirtualRegister(CopyReg))
881             SrcMI = MRI->getVRegDef(CopyReg);
882         }
883 
884         unsigned KnownZeroCount = getKnownLeadingZeroCount(SrcMI, TII);
885         if (MI.getOperand(3).getImm() <= KnownZeroCount) {
886           LLVM_DEBUG(dbgs() << "Removing redundant zero-extension\n");
887           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
888                   MI.getOperand(0).getReg())
889               .addReg(SrcReg);
890           ToErase = &MI;
891           Simplified = true;
892           NumEliminatedZExt++;
893         }
894         break;
895       }
896 
897       // TODO: Any instruction that has an immediate form fed only by a PHI
898       // whose operands are all load immediate can be folded away. We currently
899       // do this for ADD instructions, but should expand it to arithmetic and
900       // binary instructions with immediate forms in the future.
901       case PPC::ADD4:
902       case PPC::ADD8: {
903         auto isSingleUsePHI = [&](MachineOperand *PhiOp) {
904           assert(PhiOp && "Invalid Operand!");
905           MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI);
906 
907           return DefPhiMI && (DefPhiMI->getOpcode() == PPC::PHI) &&
908                  MRI->hasOneNonDBGUse(DefPhiMI->getOperand(0).getReg());
909         };
910 
911         auto dominatesAllSingleUseLIs = [&](MachineOperand *DominatorOp,
912                                             MachineOperand *PhiOp) {
913           assert(PhiOp && "Invalid Operand!");
914           assert(DominatorOp && "Invalid Operand!");
915           MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI);
916           MachineInstr *DefDomMI = getVRegDefOrNull(DominatorOp, MRI);
917 
918           // Note: the vregs only show up at odd indices position of PHI Node,
919           // the even indices position save the BB info.
920           for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) {
921             MachineInstr *LiMI =
922                 getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI);
923             if (!LiMI ||
924                 (LiMI->getOpcode() != PPC::LI && LiMI->getOpcode() != PPC::LI8)
925                 || !MRI->hasOneNonDBGUse(LiMI->getOperand(0).getReg()) ||
926                 !MDT->dominates(DefDomMI, LiMI))
927               return false;
928           }
929 
930           return true;
931         };
932 
933         MachineOperand Op1 = MI.getOperand(1);
934         MachineOperand Op2 = MI.getOperand(2);
935         if (isSingleUsePHI(&Op2) && dominatesAllSingleUseLIs(&Op1, &Op2))
936           std::swap(Op1, Op2);
937         else if (!isSingleUsePHI(&Op1) || !dominatesAllSingleUseLIs(&Op2, &Op1))
938           break; // We don't have an ADD fed by LI's that can be transformed
939 
940         // Now we know that Op1 is the PHI node and Op2 is the dominator
941         Register DominatorReg = Op2.getReg();
942 
943         const TargetRegisterClass *TRC = MI.getOpcode() == PPC::ADD8
944                                              ? &PPC::G8RC_and_G8RC_NOX0RegClass
945                                              : &PPC::GPRC_and_GPRC_NOR0RegClass;
946         MRI->setRegClass(DominatorReg, TRC);
947 
948         // replace LIs with ADDIs
949         MachineInstr *DefPhiMI = getVRegDefOrNull(&Op1, MRI);
950         for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) {
951           MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI);
952           LLVM_DEBUG(dbgs() << "Optimizing LI to ADDI: ");
953           LLVM_DEBUG(LiMI->dump());
954 
955           // There could be repeated registers in the PHI, e.g: %1 =
956           // PHI %6, <%bb.2>, %8, <%bb.3>, %8, <%bb.6>; So if we've
957           // already replaced the def instruction, skip.
958           if (LiMI->getOpcode() == PPC::ADDI || LiMI->getOpcode() == PPC::ADDI8)
959             continue;
960 
961           assert((LiMI->getOpcode() == PPC::LI ||
962                   LiMI->getOpcode() == PPC::LI8) &&
963                  "Invalid Opcode!");
964           auto LiImm = LiMI->getOperand(1).getImm(); // save the imm of LI
965           LiMI->RemoveOperand(1);                    // remove the imm of LI
966           LiMI->setDesc(TII->get(LiMI->getOpcode() == PPC::LI ? PPC::ADDI
967                                                               : PPC::ADDI8));
968           MachineInstrBuilder(*LiMI->getParent()->getParent(), *LiMI)
969               .addReg(DominatorReg)
970               .addImm(LiImm); // restore the imm of LI
971           LLVM_DEBUG(LiMI->dump());
972         }
973 
974         // Replace ADD with COPY
975         LLVM_DEBUG(dbgs() << "Optimizing ADD to COPY: ");
976         LLVM_DEBUG(MI.dump());
977         BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
978                 MI.getOperand(0).getReg())
979             .add(Op1);
980         ToErase = &MI;
981         Simplified = true;
982         NumOptADDLIs++;
983         break;
984       }
985       case PPC::RLDICR: {
986         Simplified |= emitRLDICWhenLoweringJumpTables(MI) ||
987                       combineSEXTAndSHL(MI, ToErase);
988         break;
989       }
990       case PPC::RLWINM:
991       case PPC::RLWINM_rec:
992       case PPC::RLWINM8:
993       case PPC::RLWINM8_rec: {
994         Simplified = TII->combineRLWINM(MI, &ToErase);
995         if (Simplified)
996           ++NumRotatesCollapsed;
997         break;
998       }
999       }
1000     }
1001 
1002     // If the last instruction was marked for elimination,
1003     // remove it now.
1004     if (ToErase) {
1005       ToErase->eraseFromParent();
1006       ToErase = nullptr;
1007     }
1008   }
1009 
1010   // Eliminate all the TOC save instructions which are redundant.
1011   Simplified |= eliminateRedundantTOCSaves(TOCSaves);
1012   PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>();
1013   if (FI->mustSaveTOC())
1014     NumTOCSavesInPrologue++;
1015 
1016   // We try to eliminate redundant compare instruction.
1017   Simplified |= eliminateRedundantCompare();
1018 
1019   return Simplified;
1020 }
1021 
1022 // helper functions for eliminateRedundantCompare
isEqOrNe(MachineInstr * BI)1023 static bool isEqOrNe(MachineInstr *BI) {
1024   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1025   unsigned PredCond = PPC::getPredicateCondition(Pred);
1026   return (PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE);
1027 }
1028 
isSupportedCmpOp(unsigned opCode)1029 static bool isSupportedCmpOp(unsigned opCode) {
1030   return (opCode == PPC::CMPLD  || opCode == PPC::CMPD  ||
1031           opCode == PPC::CMPLW  || opCode == PPC::CMPW  ||
1032           opCode == PPC::CMPLDI || opCode == PPC::CMPDI ||
1033           opCode == PPC::CMPLWI || opCode == PPC::CMPWI);
1034 }
1035 
is64bitCmpOp(unsigned opCode)1036 static bool is64bitCmpOp(unsigned opCode) {
1037   return (opCode == PPC::CMPLD  || opCode == PPC::CMPD ||
1038           opCode == PPC::CMPLDI || opCode == PPC::CMPDI);
1039 }
1040 
isSignedCmpOp(unsigned opCode)1041 static bool isSignedCmpOp(unsigned opCode) {
1042   return (opCode == PPC::CMPD  || opCode == PPC::CMPW ||
1043           opCode == PPC::CMPDI || opCode == PPC::CMPWI);
1044 }
1045 
getSignedCmpOpCode(unsigned opCode)1046 static unsigned getSignedCmpOpCode(unsigned opCode) {
1047   if (opCode == PPC::CMPLD)  return PPC::CMPD;
1048   if (opCode == PPC::CMPLW)  return PPC::CMPW;
1049   if (opCode == PPC::CMPLDI) return PPC::CMPDI;
1050   if (opCode == PPC::CMPLWI) return PPC::CMPWI;
1051   return opCode;
1052 }
1053 
1054 // We can decrement immediate x in (GE x) by changing it to (GT x-1) or
1055 // (LT x) to (LE x-1)
getPredicateToDecImm(MachineInstr * BI,MachineInstr * CMPI)1056 static unsigned getPredicateToDecImm(MachineInstr *BI, MachineInstr *CMPI) {
1057   uint64_t Imm = CMPI->getOperand(2).getImm();
1058   bool SignedCmp = isSignedCmpOp(CMPI->getOpcode());
1059   if ((!SignedCmp && Imm == 0) || (SignedCmp && Imm == 0x8000))
1060     return 0;
1061 
1062   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1063   unsigned PredCond = PPC::getPredicateCondition(Pred);
1064   unsigned PredHint = PPC::getPredicateHint(Pred);
1065   if (PredCond == PPC::PRED_GE)
1066     return PPC::getPredicate(PPC::PRED_GT, PredHint);
1067   if (PredCond == PPC::PRED_LT)
1068     return PPC::getPredicate(PPC::PRED_LE, PredHint);
1069 
1070   return 0;
1071 }
1072 
1073 // We can increment immediate x in (GT x) by changing it to (GE x+1) or
1074 // (LE x) to (LT x+1)
getPredicateToIncImm(MachineInstr * BI,MachineInstr * CMPI)1075 static unsigned getPredicateToIncImm(MachineInstr *BI, MachineInstr *CMPI) {
1076   uint64_t Imm = CMPI->getOperand(2).getImm();
1077   bool SignedCmp = isSignedCmpOp(CMPI->getOpcode());
1078   if ((!SignedCmp && Imm == 0xFFFF) || (SignedCmp && Imm == 0x7FFF))
1079     return 0;
1080 
1081   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1082   unsigned PredCond = PPC::getPredicateCondition(Pred);
1083   unsigned PredHint = PPC::getPredicateHint(Pred);
1084   if (PredCond == PPC::PRED_GT)
1085     return PPC::getPredicate(PPC::PRED_GE, PredHint);
1086   if (PredCond == PPC::PRED_LE)
1087     return PPC::getPredicate(PPC::PRED_LT, PredHint);
1088 
1089   return 0;
1090 }
1091 
1092 // This takes a Phi node and returns a register value for the specified BB.
getIncomingRegForBlock(MachineInstr * Phi,MachineBasicBlock * MBB)1093 static unsigned getIncomingRegForBlock(MachineInstr *Phi,
1094                                        MachineBasicBlock *MBB) {
1095   for (unsigned I = 2, E = Phi->getNumOperands() + 1; I != E; I += 2) {
1096     MachineOperand &MO = Phi->getOperand(I);
1097     if (MO.getMBB() == MBB)
1098       return Phi->getOperand(I-1).getReg();
1099   }
1100   llvm_unreachable("invalid src basic block for this Phi node\n");
1101   return 0;
1102 }
1103 
1104 // This function tracks the source of the register through register copy.
1105 // If BB1 and BB2 are non-NULL, we also track PHI instruction in BB2
1106 // assuming that the control comes from BB1 into BB2.
getSrcVReg(unsigned Reg,MachineBasicBlock * BB1,MachineBasicBlock * BB2,MachineRegisterInfo * MRI)1107 static unsigned getSrcVReg(unsigned Reg, MachineBasicBlock *BB1,
1108                            MachineBasicBlock *BB2, MachineRegisterInfo *MRI) {
1109   unsigned SrcReg = Reg;
1110   while (1) {
1111     unsigned NextReg = SrcReg;
1112     MachineInstr *Inst = MRI->getVRegDef(SrcReg);
1113     if (BB1 && Inst->getOpcode() == PPC::PHI && Inst->getParent() == BB2) {
1114       NextReg = getIncomingRegForBlock(Inst, BB1);
1115       // We track through PHI only once to avoid infinite loop.
1116       BB1 = nullptr;
1117     }
1118     else if (Inst->isFullCopy())
1119       NextReg = Inst->getOperand(1).getReg();
1120     if (NextReg == SrcReg || !Register::isVirtualRegister(NextReg))
1121       break;
1122     SrcReg = NextReg;
1123   }
1124   return SrcReg;
1125 }
1126 
eligibleForCompareElimination(MachineBasicBlock & MBB,MachineBasicBlock * & PredMBB,MachineBasicBlock * & MBBtoMoveCmp,MachineRegisterInfo * MRI)1127 static bool eligibleForCompareElimination(MachineBasicBlock &MBB,
1128                                           MachineBasicBlock *&PredMBB,
1129                                           MachineBasicBlock *&MBBtoMoveCmp,
1130                                           MachineRegisterInfo *MRI) {
1131 
1132   auto isEligibleBB = [&](MachineBasicBlock &BB) {
1133     auto BII = BB.getFirstInstrTerminator();
1134     // We optimize BBs ending with a conditional branch.
1135     // We check only for BCC here, not BCCLR, because BCCLR
1136     // will be formed only later in the pipeline.
1137     if (BB.succ_size() == 2 &&
1138         BII != BB.instr_end() &&
1139         (*BII).getOpcode() == PPC::BCC &&
1140         (*BII).getOperand(1).isReg()) {
1141       // We optimize only if the condition code is used only by one BCC.
1142       Register CndReg = (*BII).getOperand(1).getReg();
1143       if (!Register::isVirtualRegister(CndReg) || !MRI->hasOneNonDBGUse(CndReg))
1144         return false;
1145 
1146       MachineInstr *CMPI = MRI->getVRegDef(CndReg);
1147       // We assume compare and branch are in the same BB for ease of analysis.
1148       if (CMPI->getParent() != &BB)
1149         return false;
1150 
1151       // We skip this BB if a physical register is used in comparison.
1152       for (MachineOperand &MO : CMPI->operands())
1153         if (MO.isReg() && !Register::isVirtualRegister(MO.getReg()))
1154           return false;
1155 
1156       return true;
1157     }
1158     return false;
1159   };
1160 
1161   // If this BB has more than one successor, we can create a new BB and
1162   // move the compare instruction in the new BB.
1163   // So far, we do not move compare instruction to a BB having multiple
1164   // successors to avoid potentially increasing code size.
1165   auto isEligibleForMoveCmp = [](MachineBasicBlock &BB) {
1166     return BB.succ_size() == 1;
1167   };
1168 
1169   if (!isEligibleBB(MBB))
1170     return false;
1171 
1172   unsigned NumPredBBs = MBB.pred_size();
1173   if (NumPredBBs == 1) {
1174     MachineBasicBlock *TmpMBB = *MBB.pred_begin();
1175     if (isEligibleBB(*TmpMBB)) {
1176       PredMBB = TmpMBB;
1177       MBBtoMoveCmp = nullptr;
1178       return true;
1179     }
1180   }
1181   else if (NumPredBBs == 2) {
1182     // We check for partially redundant case.
1183     // So far, we support cases with only two predecessors
1184     // to avoid increasing the number of instructions.
1185     MachineBasicBlock::pred_iterator PI = MBB.pred_begin();
1186     MachineBasicBlock *Pred1MBB = *PI;
1187     MachineBasicBlock *Pred2MBB = *(PI+1);
1188 
1189     if (isEligibleBB(*Pred1MBB) && isEligibleForMoveCmp(*Pred2MBB)) {
1190       // We assume Pred1MBB is the BB containing the compare to be merged and
1191       // Pred2MBB is the BB to which we will append a compare instruction.
1192       // Hence we can proceed as is.
1193     }
1194     else if (isEligibleBB(*Pred2MBB) && isEligibleForMoveCmp(*Pred1MBB)) {
1195       // We need to swap Pred1MBB and Pred2MBB to canonicalize.
1196       std::swap(Pred1MBB, Pred2MBB);
1197     }
1198     else return false;
1199 
1200     // Here, Pred2MBB is the BB to which we need to append a compare inst.
1201     // We cannot move the compare instruction if operands are not available
1202     // in Pred2MBB (i.e. defined in MBB by an instruction other than PHI).
1203     MachineInstr *BI = &*MBB.getFirstInstrTerminator();
1204     MachineInstr *CMPI = MRI->getVRegDef(BI->getOperand(1).getReg());
1205     for (int I = 1; I <= 2; I++)
1206       if (CMPI->getOperand(I).isReg()) {
1207         MachineInstr *Inst = MRI->getVRegDef(CMPI->getOperand(I).getReg());
1208         if (Inst->getParent() == &MBB && Inst->getOpcode() != PPC::PHI)
1209           return false;
1210       }
1211 
1212     PredMBB = Pred1MBB;
1213     MBBtoMoveCmp = Pred2MBB;
1214     return true;
1215   }
1216 
1217   return false;
1218 }
1219 
1220 // This function will iterate over the input map containing a pair of TOC save
1221 // instruction and a flag. The flag will be set to false if the TOC save is
1222 // proven redundant. This function will erase from the basic block all the TOC
1223 // saves marked as redundant.
eliminateRedundantTOCSaves(std::map<MachineInstr *,bool> & TOCSaves)1224 bool PPCMIPeephole::eliminateRedundantTOCSaves(
1225     std::map<MachineInstr *, bool> &TOCSaves) {
1226   bool Simplified = false;
1227   int NumKept = 0;
1228   for (auto TOCSave : TOCSaves) {
1229     if (!TOCSave.second) {
1230       TOCSave.first->eraseFromParent();
1231       RemoveTOCSave++;
1232       Simplified = true;
1233     } else {
1234       NumKept++;
1235     }
1236   }
1237 
1238   if (NumKept > 1)
1239     MultiTOCSaves++;
1240 
1241   return Simplified;
1242 }
1243 
1244 // If multiple conditional branches are executed based on the (essentially)
1245 // same comparison, we merge compare instructions into one and make multiple
1246 // conditional branches on this comparison.
1247 // For example,
1248 //   if (a == 0) { ... }
1249 //   else if (a < 0) { ... }
1250 // can be executed by one compare and two conditional branches instead of
1251 // two pairs of a compare and a conditional branch.
1252 //
1253 // This method merges two compare instructions in two MBBs and modifies the
1254 // compare and conditional branch instructions if needed.
1255 // For the above example, the input for this pass looks like:
1256 //   cmplwi r3, 0
1257 //   beq    0, .LBB0_3
1258 //   cmpwi  r3, -1
1259 //   bgt    0, .LBB0_4
1260 // So, before merging two compares, we need to modify these instructions as
1261 //   cmpwi  r3, 0       ; cmplwi and cmpwi yield same result for beq
1262 //   beq    0, .LBB0_3
1263 //   cmpwi  r3, 0       ; greather than -1 means greater or equal to 0
1264 //   bge    0, .LBB0_4
1265 
eliminateRedundantCompare(void)1266 bool PPCMIPeephole::eliminateRedundantCompare(void) {
1267   bool Simplified = false;
1268 
1269   for (MachineBasicBlock &MBB2 : *MF) {
1270     MachineBasicBlock *MBB1 = nullptr, *MBBtoMoveCmp = nullptr;
1271 
1272     // For fully redundant case, we select two basic blocks MBB1 and MBB2
1273     // as an optimization target if
1274     // - both MBBs end with a conditional branch,
1275     // - MBB1 is the only predecessor of MBB2, and
1276     // - compare does not take a physical register as a operand in both MBBs.
1277     // In this case, eligibleForCompareElimination sets MBBtoMoveCmp nullptr.
1278     //
1279     // As partially redundant case, we additionally handle if MBB2 has one
1280     // additional predecessor, which has only one successor (MBB2).
1281     // In this case, we move the compare instruction originally in MBB2 into
1282     // MBBtoMoveCmp. This partially redundant case is typically appear by
1283     // compiling a while loop; here, MBBtoMoveCmp is the loop preheader.
1284     //
1285     // Overview of CFG of related basic blocks
1286     // Fully redundant case        Partially redundant case
1287     //   --------                   ----------------  --------
1288     //   | MBB1 | (w/ 2 succ)       | MBBtoMoveCmp |  | MBB1 | (w/ 2 succ)
1289     //   --------                   ----------------  --------
1290     //      |    \                     (w/ 1 succ) \     |    \
1291     //      |     \                                 \    |     \
1292     //      |                                        \   |
1293     //   --------                                     --------
1294     //   | MBB2 | (w/ 1 pred                          | MBB2 | (w/ 2 pred
1295     //   -------- and 2 succ)                         -------- and 2 succ)
1296     //      |    \                                       |    \
1297     //      |     \                                      |     \
1298     //
1299     if (!eligibleForCompareElimination(MBB2, MBB1, MBBtoMoveCmp, MRI))
1300       continue;
1301 
1302     MachineInstr *BI1   = &*MBB1->getFirstInstrTerminator();
1303     MachineInstr *CMPI1 = MRI->getVRegDef(BI1->getOperand(1).getReg());
1304 
1305     MachineInstr *BI2   = &*MBB2.getFirstInstrTerminator();
1306     MachineInstr *CMPI2 = MRI->getVRegDef(BI2->getOperand(1).getReg());
1307     bool IsPartiallyRedundant = (MBBtoMoveCmp != nullptr);
1308 
1309     // We cannot optimize an unsupported compare opcode or
1310     // a mix of 32-bit and 64-bit comaprisons
1311     if (!isSupportedCmpOp(CMPI1->getOpcode()) ||
1312         !isSupportedCmpOp(CMPI2->getOpcode()) ||
1313         is64bitCmpOp(CMPI1->getOpcode()) != is64bitCmpOp(CMPI2->getOpcode()))
1314       continue;
1315 
1316     unsigned NewOpCode = 0;
1317     unsigned NewPredicate1 = 0, NewPredicate2 = 0;
1318     int16_t Imm1 = 0, NewImm1 = 0, Imm2 = 0, NewImm2 = 0;
1319     bool SwapOperands = false;
1320 
1321     if (CMPI1->getOpcode() != CMPI2->getOpcode()) {
1322       // Typically, unsigned comparison is used for equality check, but
1323       // we replace it with a signed comparison if the comparison
1324       // to be merged is a signed comparison.
1325       // In other cases of opcode mismatch, we cannot optimize this.
1326 
1327       // We cannot change opcode when comparing against an immediate
1328       // if the most significant bit of the immediate is one
1329       // due to the difference in sign extension.
1330       auto CmpAgainstImmWithSignBit = [](MachineInstr *I) {
1331         if (!I->getOperand(2).isImm())
1332           return false;
1333         int16_t Imm = (int16_t)I->getOperand(2).getImm();
1334         return Imm < 0;
1335       };
1336 
1337       if (isEqOrNe(BI2) && !CmpAgainstImmWithSignBit(CMPI2) &&
1338           CMPI1->getOpcode() == getSignedCmpOpCode(CMPI2->getOpcode()))
1339         NewOpCode = CMPI1->getOpcode();
1340       else if (isEqOrNe(BI1) && !CmpAgainstImmWithSignBit(CMPI1) &&
1341                getSignedCmpOpCode(CMPI1->getOpcode()) == CMPI2->getOpcode())
1342         NewOpCode = CMPI2->getOpcode();
1343       else continue;
1344     }
1345 
1346     if (CMPI1->getOperand(2).isReg() && CMPI2->getOperand(2).isReg()) {
1347       // In case of comparisons between two registers, these two registers
1348       // must be same to merge two comparisons.
1349       unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(),
1350                                          nullptr, nullptr, MRI);
1351       unsigned Cmp1Operand2 = getSrcVReg(CMPI1->getOperand(2).getReg(),
1352                                          nullptr, nullptr, MRI);
1353       unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(),
1354                                          MBB1, &MBB2, MRI);
1355       unsigned Cmp2Operand2 = getSrcVReg(CMPI2->getOperand(2).getReg(),
1356                                          MBB1, &MBB2, MRI);
1357 
1358       if (Cmp1Operand1 == Cmp2Operand1 && Cmp1Operand2 == Cmp2Operand2) {
1359         // Same pair of registers in the same order; ready to merge as is.
1360       }
1361       else if (Cmp1Operand1 == Cmp2Operand2 && Cmp1Operand2 == Cmp2Operand1) {
1362         // Same pair of registers in different order.
1363         // We reverse the predicate to merge compare instructions.
1364         PPC::Predicate Pred = (PPC::Predicate)BI2->getOperand(0).getImm();
1365         NewPredicate2 = (unsigned)PPC::getSwappedPredicate(Pred);
1366         // In case of partial redundancy, we need to swap operands
1367         // in another compare instruction.
1368         SwapOperands = true;
1369       }
1370       else continue;
1371     }
1372     else if (CMPI1->getOperand(2).isImm() && CMPI2->getOperand(2).isImm()) {
1373       // In case of comparisons between a register and an immediate,
1374       // the operand register must be same for two compare instructions.
1375       unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(),
1376                                          nullptr, nullptr, MRI);
1377       unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(),
1378                                          MBB1, &MBB2, MRI);
1379       if (Cmp1Operand1 != Cmp2Operand1)
1380         continue;
1381 
1382       NewImm1 = Imm1 = (int16_t)CMPI1->getOperand(2).getImm();
1383       NewImm2 = Imm2 = (int16_t)CMPI2->getOperand(2).getImm();
1384 
1385       // If immediate are not same, we try to adjust by changing predicate;
1386       // e.g. GT imm means GE (imm+1).
1387       if (Imm1 != Imm2 && (!isEqOrNe(BI2) || !isEqOrNe(BI1))) {
1388         int Diff = Imm1 - Imm2;
1389         if (Diff < -2 || Diff > 2)
1390           continue;
1391 
1392         unsigned PredToInc1 = getPredicateToIncImm(BI1, CMPI1);
1393         unsigned PredToDec1 = getPredicateToDecImm(BI1, CMPI1);
1394         unsigned PredToInc2 = getPredicateToIncImm(BI2, CMPI2);
1395         unsigned PredToDec2 = getPredicateToDecImm(BI2, CMPI2);
1396         if (Diff == 2) {
1397           if (PredToInc2 && PredToDec1) {
1398             NewPredicate2 = PredToInc2;
1399             NewPredicate1 = PredToDec1;
1400             NewImm2++;
1401             NewImm1--;
1402           }
1403         }
1404         else if (Diff == 1) {
1405           if (PredToInc2) {
1406             NewImm2++;
1407             NewPredicate2 = PredToInc2;
1408           }
1409           else if (PredToDec1) {
1410             NewImm1--;
1411             NewPredicate1 = PredToDec1;
1412           }
1413         }
1414         else if (Diff == -1) {
1415           if (PredToDec2) {
1416             NewImm2--;
1417             NewPredicate2 = PredToDec2;
1418           }
1419           else if (PredToInc1) {
1420             NewImm1++;
1421             NewPredicate1 = PredToInc1;
1422           }
1423         }
1424         else if (Diff == -2) {
1425           if (PredToDec2 && PredToInc1) {
1426             NewPredicate2 = PredToDec2;
1427             NewPredicate1 = PredToInc1;
1428             NewImm2--;
1429             NewImm1++;
1430           }
1431         }
1432       }
1433 
1434       // We cannot merge two compares if the immediates are not same.
1435       if (NewImm2 != NewImm1)
1436         continue;
1437     }
1438 
1439     LLVM_DEBUG(dbgs() << "Optimize two pairs of compare and branch:\n");
1440     LLVM_DEBUG(CMPI1->dump());
1441     LLVM_DEBUG(BI1->dump());
1442     LLVM_DEBUG(CMPI2->dump());
1443     LLVM_DEBUG(BI2->dump());
1444 
1445     // We adjust opcode, predicates and immediate as we determined above.
1446     if (NewOpCode != 0 && NewOpCode != CMPI1->getOpcode()) {
1447       CMPI1->setDesc(TII->get(NewOpCode));
1448     }
1449     if (NewPredicate1) {
1450       BI1->getOperand(0).setImm(NewPredicate1);
1451     }
1452     if (NewPredicate2) {
1453       BI2->getOperand(0).setImm(NewPredicate2);
1454     }
1455     if (NewImm1 != Imm1) {
1456       CMPI1->getOperand(2).setImm(NewImm1);
1457     }
1458 
1459     if (IsPartiallyRedundant) {
1460       // We touch up the compare instruction in MBB2 and move it to
1461       // a previous BB to handle partially redundant case.
1462       if (SwapOperands) {
1463         Register Op1 = CMPI2->getOperand(1).getReg();
1464         Register Op2 = CMPI2->getOperand(2).getReg();
1465         CMPI2->getOperand(1).setReg(Op2);
1466         CMPI2->getOperand(2).setReg(Op1);
1467       }
1468       if (NewImm2 != Imm2)
1469         CMPI2->getOperand(2).setImm(NewImm2);
1470 
1471       for (int I = 1; I <= 2; I++) {
1472         if (CMPI2->getOperand(I).isReg()) {
1473           MachineInstr *Inst = MRI->getVRegDef(CMPI2->getOperand(I).getReg());
1474           if (Inst->getParent() != &MBB2)
1475             continue;
1476 
1477           assert(Inst->getOpcode() == PPC::PHI &&
1478                  "We cannot support if an operand comes from this BB.");
1479           unsigned SrcReg = getIncomingRegForBlock(Inst, MBBtoMoveCmp);
1480           CMPI2->getOperand(I).setReg(SrcReg);
1481         }
1482       }
1483       auto I = MachineBasicBlock::iterator(MBBtoMoveCmp->getFirstTerminator());
1484       MBBtoMoveCmp->splice(I, &MBB2, MachineBasicBlock::iterator(CMPI2));
1485 
1486       DebugLoc DL = CMPI2->getDebugLoc();
1487       Register NewVReg = MRI->createVirtualRegister(&PPC::CRRCRegClass);
1488       BuildMI(MBB2, MBB2.begin(), DL,
1489               TII->get(PPC::PHI), NewVReg)
1490         .addReg(BI1->getOperand(1).getReg()).addMBB(MBB1)
1491         .addReg(BI2->getOperand(1).getReg()).addMBB(MBBtoMoveCmp);
1492       BI2->getOperand(1).setReg(NewVReg);
1493     }
1494     else {
1495       // We finally eliminate compare instruction in MBB2.
1496       BI2->getOperand(1).setReg(BI1->getOperand(1).getReg());
1497       CMPI2->eraseFromParent();
1498     }
1499     BI2->getOperand(1).setIsKill(true);
1500     BI1->getOperand(1).setIsKill(false);
1501 
1502     LLVM_DEBUG(dbgs() << "into a compare and two branches:\n");
1503     LLVM_DEBUG(CMPI1->dump());
1504     LLVM_DEBUG(BI1->dump());
1505     LLVM_DEBUG(BI2->dump());
1506     if (IsPartiallyRedundant) {
1507       LLVM_DEBUG(dbgs() << "The following compare is moved into "
1508                         << printMBBReference(*MBBtoMoveCmp)
1509                         << " to handle partial redundancy.\n");
1510       LLVM_DEBUG(CMPI2->dump());
1511     }
1512 
1513     Simplified = true;
1514   }
1515 
1516   return Simplified;
1517 }
1518 
1519 // We miss the opportunity to emit an RLDIC when lowering jump tables
1520 // since ISEL sees only a single basic block. When selecting, the clear
1521 // and shift left will be in different blocks.
emitRLDICWhenLoweringJumpTables(MachineInstr & MI)1522 bool PPCMIPeephole::emitRLDICWhenLoweringJumpTables(MachineInstr &MI) {
1523   if (MI.getOpcode() != PPC::RLDICR)
1524     return false;
1525 
1526   Register SrcReg = MI.getOperand(1).getReg();
1527   if (!Register::isVirtualRegister(SrcReg))
1528     return false;
1529 
1530   MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
1531   if (SrcMI->getOpcode() != PPC::RLDICL)
1532     return false;
1533 
1534   MachineOperand MOpSHSrc = SrcMI->getOperand(2);
1535   MachineOperand MOpMBSrc = SrcMI->getOperand(3);
1536   MachineOperand MOpSHMI = MI.getOperand(2);
1537   MachineOperand MOpMEMI = MI.getOperand(3);
1538   if (!(MOpSHSrc.isImm() && MOpMBSrc.isImm() && MOpSHMI.isImm() &&
1539         MOpMEMI.isImm()))
1540     return false;
1541 
1542   uint64_t SHSrc = MOpSHSrc.getImm();
1543   uint64_t MBSrc = MOpMBSrc.getImm();
1544   uint64_t SHMI = MOpSHMI.getImm();
1545   uint64_t MEMI = MOpMEMI.getImm();
1546   uint64_t NewSH = SHSrc + SHMI;
1547   uint64_t NewMB = MBSrc - SHMI;
1548   if (NewMB > 63 || NewSH > 63)
1549     return false;
1550 
1551   // The bits cleared with RLDICL are [0, MBSrc).
1552   // The bits cleared with RLDICR are (MEMI, 63].
1553   // After the sequence, the bits cleared are:
1554   // [0, MBSrc-SHMI) and (MEMI, 63).
1555   //
1556   // The bits cleared with RLDIC are [0, NewMB) and (63-NewSH, 63].
1557   if ((63 - NewSH) != MEMI)
1558     return false;
1559 
1560   LLVM_DEBUG(dbgs() << "Converting pair: ");
1561   LLVM_DEBUG(SrcMI->dump());
1562   LLVM_DEBUG(MI.dump());
1563 
1564   MI.setDesc(TII->get(PPC::RLDIC));
1565   MI.getOperand(1).setReg(SrcMI->getOperand(1).getReg());
1566   MI.getOperand(2).setImm(NewSH);
1567   MI.getOperand(3).setImm(NewMB);
1568   MI.getOperand(1).setIsKill(SrcMI->getOperand(1).isKill());
1569   SrcMI->getOperand(1).setIsKill(false);
1570 
1571   LLVM_DEBUG(dbgs() << "To: ");
1572   LLVM_DEBUG(MI.dump());
1573   NumRotatesCollapsed++;
1574   // If SrcReg has no non-debug use it's safe to delete its def SrcMI.
1575   if (MRI->use_nodbg_empty(SrcReg)) {
1576     assert(!SrcMI->hasImplicitDef() &&
1577            "Not expecting an implicit def with this instr.");
1578     SrcMI->eraseFromParent();
1579   }
1580   return true;
1581 }
1582 
1583 // For case in LLVM IR
1584 // entry:
1585 //   %iconv = sext i32 %index to i64
1586 //   br i1 undef label %true, label %false
1587 // true:
1588 //   %ptr = getelementptr inbounds i32, i32* null, i64 %iconv
1589 // ...
1590 // PPCISelLowering::combineSHL fails to combine, because sext and shl are in
1591 // different BBs when conducting instruction selection. We can do a peephole
1592 // optimization to combine these two instructions into extswsli after
1593 // instruction selection.
combineSEXTAndSHL(MachineInstr & MI,MachineInstr * & ToErase)1594 bool PPCMIPeephole::combineSEXTAndSHL(MachineInstr &MI,
1595                                       MachineInstr *&ToErase) {
1596   if (MI.getOpcode() != PPC::RLDICR)
1597     return false;
1598 
1599   if (!MF->getSubtarget<PPCSubtarget>().isISA3_0())
1600     return false;
1601 
1602   assert(MI.getNumOperands() == 4 && "RLDICR should have 4 operands");
1603 
1604   MachineOperand MOpSHMI = MI.getOperand(2);
1605   MachineOperand MOpMEMI = MI.getOperand(3);
1606   if (!(MOpSHMI.isImm() && MOpMEMI.isImm()))
1607     return false;
1608 
1609   uint64_t SHMI = MOpSHMI.getImm();
1610   uint64_t MEMI = MOpMEMI.getImm();
1611   if (SHMI + MEMI != 63)
1612     return false;
1613 
1614   Register SrcReg = MI.getOperand(1).getReg();
1615   if (!Register::isVirtualRegister(SrcReg))
1616     return false;
1617 
1618   MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
1619   if (SrcMI->getOpcode() != PPC::EXTSW &&
1620       SrcMI->getOpcode() != PPC::EXTSW_32_64)
1621     return false;
1622 
1623   // If the register defined by extsw has more than one use, combination is not
1624   // needed.
1625   if (!MRI->hasOneNonDBGUse(SrcReg))
1626     return false;
1627 
1628   assert(SrcMI->getNumOperands() == 2 && "EXTSW should have 2 operands");
1629   assert(SrcMI->getOperand(1).isReg() &&
1630          "EXTSW's second operand should be a register");
1631   if (!Register::isVirtualRegister(SrcMI->getOperand(1).getReg()))
1632     return false;
1633 
1634   LLVM_DEBUG(dbgs() << "Combining pair: ");
1635   LLVM_DEBUG(SrcMI->dump());
1636   LLVM_DEBUG(MI.dump());
1637 
1638   MachineInstr *NewInstr =
1639       BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(),
1640               SrcMI->getOpcode() == PPC::EXTSW ? TII->get(PPC::EXTSWSLI)
1641                                                : TII->get(PPC::EXTSWSLI_32_64),
1642               MI.getOperand(0).getReg())
1643           .add(SrcMI->getOperand(1))
1644           .add(MOpSHMI);
1645   (void)NewInstr;
1646 
1647   LLVM_DEBUG(dbgs() << "TO: ");
1648   LLVM_DEBUG(NewInstr->dump());
1649   ++NumEXTSWAndSLDICombined;
1650   ToErase = &MI;
1651   // SrcMI, which is extsw, is of no use now, erase it.
1652   SrcMI->eraseFromParent();
1653   return true;
1654 }
1655 
1656 } // end default namespace
1657 
1658 INITIALIZE_PASS_BEGIN(PPCMIPeephole, DEBUG_TYPE,
1659                       "PowerPC MI Peephole Optimization", false, false)
1660 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1661 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
1662 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
1663 INITIALIZE_PASS_END(PPCMIPeephole, DEBUG_TYPE,
1664                     "PowerPC MI Peephole Optimization", false, false)
1665 
1666 char PPCMIPeephole::ID = 0;
1667 FunctionPass*
createPPCMIPeepholePass()1668 llvm::createPPCMIPeepholePass() { return new PPCMIPeephole(); }
1669 
1670